Shadow AI: How to Protect Your Data Without Slowing Innovation

Key Takeaways

• Shadow AI happens when employees use AI outside IT approval and oversight, potentially exposing sensitive business data.
• Blocking tools alone isn’t enough. Employees also need approved options that help them get their work done.
• AI risk depends on the data a tool handles, the systems it can access, and whether it can take actions independently.
• Managing that risk requires clear ownership, technical safeguards, and employee training, with ongoing measurement of adoption and data protection.

Employees are adopting AI faster than organizations can govern it. Here’s how CIOs and CTOs can regain visibility, reduce risk, and enable AI innovation securely.

AI tools your organization hasn’t approved may already be part of the workday. The harder question is what company data employees are sharing.

This is shadow AI: the use of AI tools or capabilities outside your organization’s approval and oversight.

Drafting a proposal with ChatGPT can lead a salesperson to paste in customer pricing. The same impulse to get help with a task can put proprietary code into a debugging assistant or a revenue spreadsheet into an AI tool for trend analysis.

Without IT oversight, you may have limited visibility into what employees share, where that data goes, or what those tools can access.

Protecting that data requires clear guidance on how employees can use AI and safeguards that support their work.

Shadow AI Is Growing Faster Than Governance

• 32% of data security incidents reported by organizations surveyed in Microsoft’s 2026 Data Security Index involved generative AI tools.
• 66% of workers who had used AI in PagerDuty’s 2026 survey said they’d done so even though they believed company policy prohibited their use.
• 34% of respondents in the same survey had entered customer information into public AI tools.
• 60% of generative AI users observed in Netskope’s 2025 research were using personal, unmanaged applications.

The studies measure different things, but together they show that unauthorized AI remains a concern, even as adoption of organized-managed accounts increases.

Why Blocking AI Alone Isn’t Enough

Blocking AI tools on company devices and networks doesn’t address every way employees might access them. Some may turn to personal phones or accounts to keep using AI in their work.

Samsung’s experience illustrates the challenge. After the company lifted an internal ban on ChatGPT in March 2023, employees reportedly entered confidential source code and an internal meeting transcript into the service. Samsung responded by restricting generative AI on company-owned devices and internal networks. It also instructed employees using personal devices not to submit company-related information or personal data, warning that violations could lead to disciplinary action.

That distinction matters. Restricting access on company systems and governing what employees share through personal accounts are different tasks. The Samsung incident doesn’t establish whether employees later bypassed the restrictions, but it illustrates why an AI policy needs to address both access and data handling.

A more sustainable approach pairs targeted restrictions with approved tools employees actually want to use. The question shifts from which applications to block to which activities to permit, under what conditions, and with what safeguards.

The Real Risk Isn’t Just the AI Tool. It’s What the Tool Can Access.

Using an approved AI assistant to brainstorm presentation ideas is fundamentally different from uploading confidential financial records or allowing an autonomous agent to modify information in a customer database.

Evaluate AI risk across three dimensions:

1. Data sensitivity. What information can the application process? Does it include customer records, intellectual property, financial information, or regulated data?
2. Application access. Is the tool approved and managed? Can the organization enforce identity, access, data handling, and security policies?
3. Level of autonomy. Does the AI simply generate recommendations, or can it retrieve information, interact with business systems, and perform actions independently?

Netskope’s 2025 research found growing adoption of generative AI platforms and custom agents, extending shadow AI beyond consumer chatbots. This makes an AI tool’s ability to access business systems and act independently an important part of the risk assessment.

An assistant that summarizes documents can expose what it reads. An agent that can send email, pull customer records, or change business data can act on that information, often before anyone reviews it.

Five Ways to Regain Control Without Slowing Innovation

1. Discover AI Usage Before Making Decisions

Start by identifying AI tools employees access, including public chatbots, coding assistants, browser extensions, AI features embedded in SaaS applications, and connected agents. Cloud application discovery and security monitoring will show you usage across managed devices and networks.

Then review past activity. Where available, historical proxy, Cloud Access Security Broker (CASB), and Data Loss Prevention (DLP) logs can help identify sensitive information shared with AI tools. What you can establish will depend on the logging and inspection coverage in place at the time. Review the relevant vendors’ retention and model-training terms, and request deletion where contracts allow. Treat confirmed exposure of regulated or customer data as an incident, not a policy reminder.

But discovery should be more than an inventory exercise. Assess which applications process sensitive information, which departments rely on them, and which represent meaningful business risks. Aim to have that picture within 30 days, with the tools that have already handled customer or regulated data flagged for follow-up.

2. Make Approved AI Easier to Use Than Shadow AI

Employees often adopt unauthorized AI because available tools don’t meet their needs or because the approval process is too complicated.

Work with business leaders to identify legitimate use cases and evaluate enterprise AI alternatives that provide appropriate security, administrative controls, and contractual protections. At minimum, confirm vendors won’t train models on your data, define retention and deletion terms, and check obligations under the regulations that apply to you, such as GDPR, HIPAA, or the EU AI Act.

Make the approved tools accessible, communicate their capabilities, and establish a clear process for requesting new applications.

Rather than treating every AI application as equally risky, prioritize reviews based on the information it processes and the business activities it supports. Start with the use cases discovery turned up most often, and commit to answering new requests in days, not months. If approval takes a quarter, employees won’t wait for it.

3. Assign Ownership and Establish Clear Governance

A policy without owners is just a document. Decide who approves new AI applications, assesses data protection requirements, accepts residual risk, and oversees AI integrations.

Policies should address more than approved applications. They should define acceptable data use, employee responsibilities, vendor evaluation requirements, and the conditions under which AI agents can access or modify business systems. Make one person ultimately accountable, and write the policy so someone in sales, finance, or engineering can tell at a glance what data they can and can’t put into an AI tool.

4. Apply Security Controls to Applications and Data

AI governance needs technical enforcement.

CASB and secure web gateway capabilities identify AI applications, classify them by risk, and enforce access policies.

Generative AI DLP adds another layer by detecting and blocking sensitive information shared through monitored AI prompts and file uploads.

Each control answers a different question:

Review coverage in four areas that may require additional configuration or controls beyond your existing CASB and DLP deployment:

• Embedded SaaS AI: AI features in Microsoft 365, Salesforce, Notion, and similar platforms run inside applications you’ve already approved. Review each vendor’s AI settings, data-handling terms, and admin controls, and disable features you haven’t evaluated.
• Browser extensions: AI extensions can read page content, including internal applications. Manage extensions through browser policy and allow only reviewed ones.
• OAuth-connected apps and agents: Employees can grant AI tools standing access to email, files, and calendars in a few clicks. Audit OAuth grants in your identity provider, restrict user consent, and revoke unused access.
• Personal devices: Review what employees can access from unmanaged phones and other personal devices. Depending on the application and configuration, conditional access and session controls can block or limit access and restrict activities such as downloading files. Pair those controls with approved AI tools that meet employees’ work needs.

Treat AI agents like privileged service accounts: least-privilege permissions, scoped and expiring credentials, human approval for actions that change records or send external communications, and a log of every action taken.

How far you go with each control should depend on what discovery turned up.

5. Train Employees and Measure What Matters

AI governance shouldn’t be a one-time initiative, and training gaps are part of the problem. A 2025 SAP WalkMe survey found that 51% of workers received conflicting guidance about AI use, while only 7.5% reported receiving extensive training.

Provide training that reflects employees’ actual responsibilities. Developers need guidance on sharing source code; sales and finance teams need examples involving customer and financial data.

Then measure whether it’s working. Three indicators matter most:

• AI application visibility: The percentage of discovered AI applications that have been reviewed and classified.
• Approved AI adoption: The percentage of AI users accessing sanctioned tools rather than unmanaged alternatives, where usage can be measured.
• Data protection trends: Changes in sensitive-data exposure alerts, policy violations, and confirmed incidents.

Interpret these measures together. An increase in alerts, for example, may reflect improved detection rather than deteriorating security.

Report these metrics to the executive team quarterly, alongside productivity results from approved AI tools, so security and adoption are judged together.

From Shadow AI to Secure Innovation

Shadow AI is a demand signal. Employees have already decided AI is worth using; the only question is whether they use it where you can see it.

You won’t eliminate unauthorized AI overnight. The goal is to make the secure path the easiest one, and to shrink your exposure every quarter.

How TPx Can Help

The controls in step 4 only work if they’re enforced consistently across users, devices, and applications, which is hard to sustain with a stretched security team.

TPx Managed Secure Service Edge (SSE) combines secure web gateways, CASB, and DLP to help organizations identify AI application usage, manage access to approved services, and protect sensitive data shared through supported AI applications.

TPx also offers cybersecurity assessments and advisory services to help organizations identify security gaps and prioritize improvements.

Talk with a TPx expert about your approach to AI security.

Sources

1. PagerDuty, Shadow AI Survey, June 11, 2026.
2. Microsoft, 2026 Data Security Index, January 29, 2026.
3. Netskope Threat Labs, Cloud and Threat Report: Shadow AI and Agentic AI 2025, 2025.
4. SAP WalkMe, AI in the Workplace Survey, August 27, 2025.
5. Forbes, Samsung Bans ChatGPT Among Employees After Sensitive Code Leak, May 2, 2023.